src/freedreno/ir3/ir3_nir_lower_io_offsets.c

   1 /*
   2  * Copyright © 2018-2019 Igalia S.L.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include "ir3_nir.h"
  25 #include "compiler/nir/nir_builder.h"
  26
  27 /**
  28  * This pass moves to NIR certain offset computations for different I/O
  29  * ops that are currently implemented on the IR3 backend compiler, to
  30  * give NIR a chance to optimize them:
  31  *
  32  * - Dword-offset for SSBO load, store and atomics: A new, similar intrinsic
  33  *   is emitted that replaces the original one, adding a new source that
  34  *   holds the result of the original byte-offset source divided by 4.
  35  */
  36
  37
  38 /* Returns the ir3-specific intrinsic opcode corresponding to an SSBO
  39  * instruction that is handled by this pass. It also conveniently returns
  40  * the offset source index in @offset_src_idx.
  41  *
  42  * If @intrinsic is not SSBO, or it is not handled by the pass, -1 is
  43  * returned.
  44  */
  45 static int
  46 get_ir3_intrinsic_for_ssbo_intrinsic(unsigned intrinsic,
  47                                                                          uint8_t *offset_src_idx)
  48 {
  49         debug_assert(offset_src_idx);
  50
  51         *offset_src_idx = 1;
  52
  53         switch (intrinsic) {
  54         case nir_intrinsic_store_ssbo:
  55                 *offset_src_idx = 2;
  56                 return nir_intrinsic_store_ssbo_ir3;
  57         case nir_intrinsic_load_ssbo:
  58                 return nir_intrinsic_load_ssbo_ir3;
  59         case nir_intrinsic_ssbo_atomic_add:
  60                 return nir_intrinsic_ssbo_atomic_add_ir3;
  61         case nir_intrinsic_ssbo_atomic_imin:
  62                 return nir_intrinsic_ssbo_atomic_imin_ir3;
  63         case nir_intrinsic_ssbo_atomic_umin:
  64                 return nir_intrinsic_ssbo_atomic_umin_ir3;
  65         case nir_intrinsic_ssbo_atomic_imax:
  66                 return nir_intrinsic_ssbo_atomic_imax_ir3;
  67         case nir_intrinsic_ssbo_atomic_umax:
  68                 return nir_intrinsic_ssbo_atomic_umax_ir3;
  69         case nir_intrinsic_ssbo_atomic_and:
  70                 return nir_intrinsic_ssbo_atomic_and_ir3;
  71         case nir_intrinsic_ssbo_atomic_or:
  72                 return nir_intrinsic_ssbo_atomic_or_ir3;
  73         case nir_intrinsic_ssbo_atomic_xor:
  74                 return nir_intrinsic_ssbo_atomic_xor_ir3;
  75         case nir_intrinsic_ssbo_atomic_exchange:
  76                 return nir_intrinsic_ssbo_atomic_exchange_ir3;
  77         case nir_intrinsic_ssbo_atomic_comp_swap:
  78                 return nir_intrinsic_ssbo_atomic_comp_swap_ir3;
  79         default:
  80                 break;
  81         }
  82
  83         return -1;
  84 }
  85
  86 static nir_ssa_def *
  87 check_and_propagate_bit_shift32(nir_builder *b, nir_alu_instr *alu_instr,
  88                                                                 int32_t direction, int32_t shift)
  89 {
  90         debug_assert(alu_instr->src[1].src.is_ssa);
  91         nir_ssa_def *shift_ssa = alu_instr->src[1].src.ssa;
  92
  93         /* Only propagate if the shift is a const value so we can check value range
  94          * statically.
  95          */
  96         nir_const_value *const_val = nir_src_as_const_value(alu_instr->src[1].src);
  97         if (!const_val)
  98                 return NULL;
  99
 100         int32_t current_shift = const_val[0].i32 * direction;
 101         int32_t new_shift = current_shift + shift;
 102
 103         /* If the merge would reverse the direction, bail out.
 104          * e.g, 'x << 2' then 'x >> 4' is not 'x >> 2'.
 105          */
 106         if (current_shift * new_shift < 0)
 107                 return NULL;
 108
 109         /* If the propagation would overflow an int32_t, bail out too to be on the
 110          * safe side.
 111          */
 112         if (new_shift < -31 || new_shift > 31)
 113                 return NULL;
 114
 115         /* Add or substract shift depending on the final direction (SHR vs. SHL). */
 116         if (shift * direction < 0)
 117                 shift_ssa = nir_isub(b, shift_ssa, nir_imm_int(b, abs(shift)));
 118         else
 119                 shift_ssa = nir_iadd(b, shift_ssa, nir_imm_int(b, abs(shift)));
 120
 121         return shift_ssa;
 122 }
 123
 124 nir_ssa_def *
 125 ir3_nir_try_propagate_bit_shift(nir_builder *b, nir_ssa_def *offset, int32_t shift)
 126 {
 127         nir_instr *offset_instr = offset->parent_instr;
 128         if (offset_instr->type != nir_instr_type_alu)
 129                 return NULL;
 130
 131         nir_alu_instr *alu = nir_instr_as_alu(offset_instr);
 132         nir_ssa_def *shift_ssa;
 133         nir_ssa_def *new_offset = NULL;
 134
 135         b->cursor = nir_after_instr(&alu->instr);
 136
 137         /* the first src could be something like ssa_18.x, but we only want
 138          * the single component.  Otherwise the ishl/ishr/ushr could turn
 139          * into a vec4 operation:
 140          */
 141         nir_ssa_def *src0 = nir_mov_alu(b, alu->src[0], 1);
 142
 143         switch (alu->op) {
 144         case nir_op_ishl:
 145                 shift_ssa = check_and_propagate_bit_shift32(b, alu, 1, shift);
 146                 if (shift_ssa)
 147                         new_offset = nir_ishl(b, src0, shift_ssa);
 148                 break;
 149         case nir_op_ishr:
 150                 shift_ssa = check_and_propagate_bit_shift32(b, alu, -1, shift);
 151                 if (shift_ssa)
 152                         new_offset = nir_ishr(b, src0, shift_ssa);
 153                 break;
 154         case nir_op_ushr:
 155                 shift_ssa = check_and_propagate_bit_shift32(b, alu, -1, shift);
 156                 if (shift_ssa)
 157                         new_offset = nir_ushr(b, src0, shift_ssa);
 158                 break;
 159         default:
 160                 return NULL;
 161         }
 162
 163         return new_offset;
 164 }
 165
 166 static bool
 167 lower_offset_for_ssbo(nir_intrinsic_instr *intrinsic, nir_builder *b,
 168                                           unsigned ir3_ssbo_opcode, uint8_t offset_src_idx)
 169 {
 170         unsigned num_srcs = nir_intrinsic_infos[intrinsic->intrinsic].num_srcs;
 171
 172         bool has_dest = nir_intrinsic_infos[intrinsic->intrinsic].has_dest;
 173         nir_ssa_def *new_dest = NULL;
 174
 175         /* Here we create a new intrinsic and copy over all contents from the old one. */
 176
 177         nir_intrinsic_instr *new_intrinsic;
 178         nir_src *target_src;
 179
 180         /* 'offset_src_idx' holds the index of the source that represent the offset. */
 181         new_intrinsic =
 182                 nir_intrinsic_instr_create(b->shader, ir3_ssbo_opcode);
 183
 184         debug_assert(intrinsic->src[offset_src_idx].is_ssa);
 185         nir_ssa_def *offset = intrinsic->src[offset_src_idx].ssa;
 186
 187         /* Since we don't have value range checking, we first try to propagate
 188          * the division by 4 ('offset >> 2') into another bit-shift instruction that
 189          * possibly defines the offset. If that's the case, we emit a similar
 190          * instructions adjusting (merging) the shift value.
 191          *
 192          * Here we use the convention that shifting right is negative while shifting
 193          * left is positive. So 'x / 4' ~ 'x >> 2' or 'x << -2'.
 194          */
 195         nir_ssa_def *new_offset = ir3_nir_try_propagate_bit_shift(b, offset, -2);
 196
 197         /* The new source that will hold the dword-offset is always the last
 198          * one for every intrinsic.
 199          */
 200         target_src = &new_intrinsic->src[num_srcs];
 201         *target_src = nir_src_for_ssa(offset);
 202
 203         if (has_dest) {
 204                 debug_assert(intrinsic->dest.is_ssa);
 205                 nir_ssa_def *dest = &intrinsic->dest.ssa;
 206                 nir_ssa_dest_init(&new_intrinsic->instr, &new_intrinsic->dest,
 207                                                   dest->num_components, dest->bit_size, NULL);
 208                 new_dest = &new_intrinsic->dest.ssa;
 209         }
 210
 211         for (unsigned i = 0; i < num_srcs; i++)
 212                 new_intrinsic->src[i] = nir_src_for_ssa(intrinsic->src[i].ssa);
 213
 214         for (unsigned i = 0; i < NIR_INTRINSIC_MAX_CONST_INDEX; i++)
 215                 new_intrinsic->const_index[i] = intrinsic->const_index[i];
 216
 217         new_intrinsic->num_components = intrinsic->num_components;
 218
 219         b->cursor = nir_before_instr(&intrinsic->instr);
 220
 221         /* If we managed to propagate the division by 4, just use the new offset
 222          * register and don't emit the SHR.
 223          */
 224         if (new_offset)
 225                 offset = new_offset;
 226         else
 227                 offset = nir_ushr(b, offset, nir_imm_int(b, 2));
 228
 229         /* Insert the new intrinsic right before the old one. */
 230         nir_builder_instr_insert(b, &new_intrinsic->instr);
 231
 232         /* Replace the last source of the new intrinsic by the result of
 233          * the offset divided by 4.
 234          */
 235         nir_instr_rewrite_src(&new_intrinsic->instr,
 236                                                   target_src,
 237                                                   nir_src_for_ssa(offset));
 238
 239         if (has_dest) {
 240                 /* Replace the uses of the original destination by that
 241                  * of the new intrinsic.
 242                  */
 243                 nir_ssa_def_rewrite_uses(&intrinsic->dest.ssa,
 244                                                                  nir_src_for_ssa(new_dest));
 245         }
 246
 247         /* Finally remove the original intrinsic. */
 248         nir_instr_remove(&intrinsic->instr);
 249
 250         return true;
 251 }
 252
 253 static bool
 254 lower_io_offsets_block(nir_block *block, nir_builder *b, void *mem_ctx)
 255 {
 256         bool progress = false;
 257
 258         nir_foreach_instr_safe(instr, block) {
 259                 if (instr->type != nir_instr_type_intrinsic)
 260                         continue;
 261
 262                 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
 263
 264                 /* SSBO */
 265                 int ir3_intrinsic;
 266                 uint8_t offset_src_idx;
 267                 ir3_intrinsic = get_ir3_intrinsic_for_ssbo_intrinsic(intr->intrinsic,
 268                                                                                                                          &offset_src_idx);
 269                 if (ir3_intrinsic != -1) {
 270                         progress |= lower_offset_for_ssbo(intr, b, (unsigned) ir3_intrinsic,
 271                                                                                           offset_src_idx);
 272                 }
 273         }
 274
 275         return progress;
 276 }
 277
 278 static bool
 279 lower_io_offsets_func(nir_function_impl *impl)
 280 {
 281         void *mem_ctx = ralloc_parent(impl);
 282         nir_builder b;
 283         nir_builder_init(&b, impl);
 284
 285         bool progress = false;
 286         nir_foreach_block_safe(block, impl) {
 287                 progress |= lower_io_offsets_block(block, &b, mem_ctx);
 288         }
 289
 290         if (progress) {
 291                 nir_metadata_preserve(impl, nir_metadata_block_index |
 292                                                                         nir_metadata_dominance);
 293         }
 294
 295         return progress;
 296 }
 297
 298 bool
 299 ir3_nir_lower_io_offsets(nir_shader *shader)
 300 {
 301         bool progress = false;
 302
 303         nir_foreach_function(function, shader) {
 304                 if (function->impl)
 305                         progress |= lower_io_offsets_func(function->impl);
 306         }
 307
 308         return progress;
 309 }